Energy Minimization Approach to Pixel Labeling Problems in Computer Vision

点击翻译按钮获取中文摘要

The goal of computer vision is to automatically analyze visual data, such as images and video. Computer vision systems are used nowadays for a wide range of applications: medical image processing, robot navigation, automotive safety, special effects in movies, image search, etc. Despite tremendous progress, there is still a large gap between what a human and a computer vision system can do. A framework that has proven very effective for a variety of vision problems is pixel labeling. It is usually addressed with energy minimization. The major difficulty of this framework is its computational cost, with many energies arising in practice being difficult to minimize. The energies that are currently minimized well are essentially those that encode low-level smoothness properties on a labeling, namely, that most adjacent pixel pairs should to have the same or similar labels. However, in a variety of vision tasks, other properties, such as preference for a particular subset of label configurations between pixel pairs, or constraints on pixel subsets of size larger than two arise. Energies with these more general constraints are still difficult to minimize. My goal is to develop efficient energy minimization algorithms for these harder-to-optimize pixel labeling problems, and to improve performance of practical applications utilizing minimization algorithms. My research is heavily based on discrete optimization algorithms, in particular on dynamic programming and graph cuts. The proposed applications for evaluating the minimization methods that I will develop are cancerous gland segmentation in prostate images for volume measuring and prognosis and shape priors for image segmentation. More generally, the algorithms that I plan to develop will lead to an improved performance for a wider range of applications, such as object recognition, medical image segmentation, motion correspondence, 3D modeling, etc.